Object Detection and 6-DoF Pose Estimation with Dense Depth Sensors using sampling-based Bayesian State Estimation

Abstract

To enable robots to execute interesting tasks and to interact with their environment the semantic gap between raw sensory data and a meaningful representation of their surroundings needs to be overcome. To his end, objects in the observed scene need to be recognized and their exact positions relative to the robot have to be detected. This work addresses the problem of identifying and estimating the 6-DoF pose (translation and orientation) of known arbitrarily shaped objects, sensed by a dense depth sensor mounted on an industrial robot manipulator. Object occlusions in complex human environments motivate an approach which includes pose estimates from previous time steps and enhances the understanding of the currently observed scene at any particular time step. We propose an object detection and 6-DoF pose estimation method using sampling-based Bayesian state estimation. For every segmented 3D data cluster in the gathered depth images, multiple hypotheses for the object identities and the corresponding 6-DoF poses are initialized. The search for the correct identities and poses is performed using particle filtering methods. The presented approach is capable of handling multiple partly visible objects and is evaluated on real depth data. It is shown that sufficient object detection rates and accurate pose estimation results are achieved to enable subsequent manipulation tasks.